Revolutionizing Access to AI Compute
Alibaba Cloud has officially launched the commercial availability of its Lingjun Zhenwu M890 supernode instances. This development allows businesses to rent high-speed, 64-card interconnected computing clusters, bypassing the need to construct and manage private data centers. This marks a significant milestone as the first Chinese-developed supernode made directly accessible via public cloud infrastructure without requiring long-term purchase agreements or pre-configured AI model setups.
However, interested users must note a geographic restriction: current access is limited exclusively to the Ulanqab region in North China's Inner Mongolia Autonomous Region.
Understanding the Supernode Architecture
A supernode functions by linking multiple accelerators so tightly that software perceives the collective as a single, massive processor rather than an array of individual units. This architecture is essential for high-end AI development, as it eliminates the performance bottlenecks caused by slow network data transfers between disparate chips. Without such tight integration, running trillion-parameter models would be impractical due to the massive overhead involved.
While industry leaders like Nvidia’s GB200 NVL72 and AMD’s 72-GPU Helios configuration represent the global state-of-the-art, Alibaba’s solution offers unique strategic advantages within China. By utilizing domestic silicon and local cloud infrastructure, Alibaba provides lower latency and a viable alternative for businesses facing restrictions on imported high-end chips.
Technical Specifications and Performance
The Zhenwu M890 processor, designed by Alibaba's T-Head unit, features 144GB of HBM memory and achieves inter-chip bandwidth speeds of up to 800 GB/s. The system utilizes the proprietary ICNSwitch 1.0, which allows for scaling up to 64 cards—a significant upgrade over the 16-card limit found in the previous Zhenwu 810E generation.
«The configuration is designed to handle Mixture of Experts (MoE) AI models of up to 10 trillion parameters, making it the first supernode architecture in China capable of efficiently executing large language models exceeding 2 trillion parameters.»
According to Alibaba, this infrastructure is already supporting commercial services for advanced models, including Kimi K3 and Qwen3.8-Max.
A Roadmap for Future Growth
Alibaba is aggressively expanding its chip capabilities. The roadmap includes the Zhenwu V900, expected in Q3 2027, which promises three times the performance of the M890, 216GB of memory, and 1,200 GB/s of bandwidth. A subsequent release, the Zhenwu J900, is slated for 2028.
While public pricing and detailed power consumption data remain unavailable, the rollout of the M890 highlights China’s commitment to self-reliance in the artificial intelligence sector. As the domestic hardware ecosystem matures, these powerful compute resources may eventually become accessible to a broader, global customer base.